-
Notifications
You must be signed in to change notification settings - Fork 4
Open
Milestone
Description
Also check for a complete list: https://github.com/rwth-i6/returnn_common/milestone/1
- Rec design for recurrent definitions / loops #16
- Design/Handling of dimension tags #17
- How to define whether search (or train flag) is enabled? #18
- Functional layer API, conventions #21
- Masked computation wrapper #23
- Cond wrapper #24
- Consistent variable name scopes for modules #25
- RETURNN layers with hidden state should make it explicit #31
Sequentialmodule #33ModuleListmodule #34Loop.Stateincomplete #35- Losses to implement and losses naming conventions #38
- Reorganization of code and module names, and user network code conventions #39
- Remove automatic layer input concatenation? #41
- Make minimum Python version explicit #43
- Better way to define main network dict (
make_root_net_dict) #44 - Replace
reduce(x, mode="mean", ...)byreduce_mean(x, ...), etc? #46 - Make shape and dims available? #47
- Enforce dim tags to be unique? #48
- Higher-level encoder decoder interfaces for transducer, attention, LM, ILM, etc #49
- Name scope for custom module functions? #50
- How to handle Sisyphus hashes #51
- Implement standard attention and self-attention module #52
- Implement standard Transformer encoder and decoder #53
- Implement standard Conformer encoder #54
- Definition of losses,
mark_as_loss? #56 - Collecting frame error, label error, edit distance? #57
- How to define the API for parameter initialization, regularization (L2, weight dropout, etc), maybe updater opts per-param #59
- Have
Conv1d,Conv2d,Conv3dinstead ofConv? Same forPool? #61 - Remove
activationoption fromLinear,Conv#62 - Make
activationfunction private, prefer all direct wrappers likereluetc #63 - Make
evalfunction private, or rename it? #64 - Implement SpecAugment #65
nn.dotshould have better signature #67- How to flatten (pack) sequence before softmax + loss or just loss #68
- Unexpected default for nn.softmax #69
- RecLayer time dim explicit instead of implicit via RecUnstackLayer #72
- Transformer example with relative positional encoding #74
- Inconsistency: input dim needed for custom modules with custom parameters but not builtin layer wrappers? #75
- att_dropout is probably applied wrongly (on the feature) #76
nn.dropoutshould haveaxismandatory #77- Change
nn.SelfAttentiondefaultatt_dropout? #78 SelfAttentionSteplacksinitial_state#79- Implement
Loop.last(or maybe only for states) #80 - Unify rec
...Stepvariants with on-seq variants? #81 - Only functional wrapped layers except
VariableLayer? #82 - Fix batch norm defaults #83
- Make sure a recent RETURNN behavior version is used? #84
- How to have custom updates for parameters #90
- How to handle parameter initialization #92
- Model checkpoint load and store logic #93
- What param init defaults should we use #94
- Use stateless random ops #95
- Parameter random init outside name ctx does not work #109
- Remove unused layers #111
name_scopelayer option not used ideally #112- Dim tag var names in serialized code are too long #113
- Loop max_seq_len support incomplete #114
- Out spatial dim tags arguments not needed #115
- Out dim tags potentially needs duplicated logic #116
- Dim tag description and identifier name inconsistent and not optimal #119
- Inconsistency for axis or in_spatial_dim #120
Diminternals and API should be refactored returnn#975- Improve verify_out_shape, ignore unrelated dims #121
- Root module, naming logic API #125
- Conformer misses relative pos encoding #132
- Broadcasting with standard operators #136
- Introduce DType class or so? #137
- constant, random, reduce: shape requires ordering #138
- Allow anonymous parameters? #147
nn.Randomfor multiple ops #148- Weight Decay #149
- Serialization of big Numpy arrays #150
- Loop not simple enough in some cases #151
- Remove
nn.scoped#159 - mark_as_loss: extension for LR scheduling? #211
- Lazy init causes unexpected behavior? #212
- Conformer/Transformer has same initial param value in each layer #216
- SelfAttention misses Linear after attention, wrong for Conformer, Transformer #221
Metadata
Metadata
Labels
No labels